How Many AI Images Does It Take to Get One You Keep?

You raise your keep rate by cutting the number of decisions a model has to make, not by generating more images. On VdoBloom that means deciding the shape and resolution before you spend, writing the prompt as a list of what must not change, matching the model to the job, and fixing a near-miss in the edit tools instead of rolling the dice again. Every generation is priced before you run it, so a shorter path to one good image is also the cheapest path.

Updated October 2026.

Why most generated images get thrown away

When an image misses, it is usually not because the model is weak. It is because the request was open. A prompt like “a woman in a red dress on a street” leaves the model free to choose the age, the framing, the light, the lens, the pose, the weather and the background, and it will choose differently every time you press the button. Nothing in that sentence tells it which of those choices matter to you.

The practical fix is to treat a prompt as a specification with three parts: the subject you want exactly, the shot you want exactly, and a short list of things that must not change. That is the whole method. The rest of this guide is the order to do it in, so you stop paying for images you already know you will discard.

Step 1: decide the shot before you look at the model list

Choose the aspect ratio first. It is the single biggest cause of a wasted generation, because a good image cropped badly is a bad image. The Generate tab exposes an aspect-ratio selector whenever the model supports one, and the lists differ: Nano Banana 2 and 2.1 offer fifteen ratios including auto and the extreme 1:8 and 8:1 crops, GPT Image 2 offers eleven including auto, and the Seedream family offers eight. Decide 1:1, 4:5 or 16:9 for the place the image is going before you generate anything.
Then choose the resolution, because it is priced. Several models charge by resolution rather than per image, and the jump is real: Nano Banana 2.1 is 3 credits at 1K, 4 at 2K and 6 at 4K, while GPT Image 2 and both GPT Image 2.5 engines are 3, 5 and 8. If the image is going into a feed, 2K is usually enough. Draft at the cheap tier and re-run the winner high, rather than discovering at 4K that the composition was wrong.
Only then pick the model. The model is the last decision because it is the one that changes least about whether the image is usable.

Step 2: write the prompt as a specification

A prompt that keeps more of what it makes has four parts, in this order:

The subject, precisely. Who or what, how many, doing what, wearing what. “A woman in her thirties in a plain red linen dress” beats “a woman” because it removes choices.
The shot. Framing, angle, lens feel, depth of field. “Three-quarter length, eye level, 85mm, shallow depth of field” tells the model it is making a portrait rather than a scene.
The light and the place. “Soft window light from the left, plain warm-grey studio wall” is worth more than any style adjective.
The exclusions. A short negative list is the highest-value line in the prompt: “no text, no logos, no extra people, no heavy retouching, natural skin texture”. It is exactly the line that stops the plastic-skin look and the invented watermark.

One habit separates people who get what they asked for from people who do not: reuse the phrasing that worked. When a prompt returns something good, keep the whole sentence and change one clause. Writing a fresh prompt each time means re-deciding everything, and the result is a different-looking image even when the subject is identical. The mechanics of this are covered in how to write a better AI image prompt.

Step 3: match the model to the job

Most waste comes from using one favourite model for everything. These are the jobs where the differences are actually visible, with the real credit cost of each run.

Job	Model to reach for	Credits per image
Fast drafts you may delete	Muse Image, or Seedream 5 Flash	1, or 2
Text and typography in the image	GPT Image 2, or GPT Image 2.5 Sunburst	3 / 5 / 8 by resolution
Photoreal people, keep the face you have	Nano Banana 2.1 (up to 14 reference images)	3 / 4 / 6 by resolution
Large prints at 4K	Nano Banana Pro, Seedream 4.5	18 / 18 / 24, or 7
Exact pixel dimensions and brand colour	Recraft V4 and V4 Pro	12, or 22
Long, detailed prompts	GPT Image 2 (up to 20,000 characters)	3 / 5 / 8 by resolution

Prompt length is a practical constraint, not a small print detail. FLUX Kontext Pro and Max, Seedream 4.5 and Seedream 5 Lite cap the box at 3,000 characters, WAN 2.7 and Seedream 5 Flash at 5,000, and Recraft V4 at 10,000. If your prompt is long because it is precise, that is a reason to pick a model with room rather than to trim the detail you need. The multi-model image generator explains the same choices from the other direction: one prompt, several models, compare before you commit.

Step 4: judge at full size, then fix instead of restarting

Open the image at full size before you decide. Composition reads differently at thumbnail scale, and the two failure modes people notice last are hands and text — which is exactly what the anatomy problem and the text problem are about.

If the image is close but not right, do not re-roll the whole thing. The edit tab works from the image you already paid for: upload it, describe the change as an instruction, and keep everything else. Most edit models take reference images — Nano Banana 2.1 and Seedream 5 Lite up to 14, GPT Image 2.5 up to 16, Flux 2 Pro up to 8 — and Qwen Image 2.1 Edit adds a painted mask for changing one area while the rest stays pixel-identical. If you do not want to pick, the Auto option is a 7-credit chain that tries Seedream 5 Flash, then Seedream 4.5, then WAN 2.7.

That is why the keep rate goes up: an edit pass on a known-good image is cheaper than another draw on an unknown one. Fixing one arm costs one edit. Re-running the prompt costs a generation and might return a different pose entirely.

Common mistakes that burn credits

Generating before choosing the ratio. The most common one. A 16:9 image cannot become a 4:5 portrait without a re-run or an outpainting pass.
Stuffing the prompt to look impressive. Three style adjectives fight each other. Pick one look and describe it properly.
Forgetting the exclusions. Without a negative list you get logo-shaped smudges, extra fingers and airbrushed skin by default.
Changing the model mid-project. The same prompt returns a different picture on a different engine, so a set built across three models looks like three sets.
Never using the edit pass. People re-generate an 80% image instead of editing the missing 20%, then wonder why nothing matches.
Not reading the refund rule. A generation that fails for a technical reason, or that the provider’s moderation blocks, is refunded automatically — so a blocked prompt costs you an attempt, not credits, and is worth rewording rather than abandoning.

Which model to pick when you do not want to think about it

If you want one safe default: Nano Banana 2.1 at 3 credits for 1K. It takes up to 14 reference images, offers fifteen aspect ratios and comes in at the cheap end of the resolution tiers, which makes it the best value for iterating towards one image you keep. Move up to Nano Banana Pro only when you genuinely need 4K, and reach for GPT Image 2 when the image has to contain readable type. Everything is priced before you press Generate, and the plans are on the pricing page, including the 10 free signup credits that cover a first round of cheap drafts.

Run the whole loop twice and the pattern is obvious: fewer decisions per image, priced correctly, fixed rather than discarded. That is a keep rate you control.

Frequently asked questions

How many AI images does it take to get one you keep?

There is no honest universal number — it depends entirely on how specific your prompt is. The practical goal is to get to one usable image in a small number of cheap drafts: use a low-cost model at 1K or 2K to settle the composition, then re-run the winner at high resolution or fix it with an edit pass.

Is it cheaper to generate several images at once or one at a time?

You pay per image, so generating more never costs less per usable result. On VdoBloom the Generate tab produces one image per run, and the price of the run is shown on the button before you spend anything. Iterating on the prompt is the cheap lever; volume is not.

What does a generation cost if the image comes out badly?

A completed generation is charged, whether or not you like the result. Two cases are different: if the job fails on the technical side, or if the provider’s content moderation blocks it, the credits are refunded automatically and you can reword and retry.

Should I fix a bad image or generate a new one?

Fix it, if the composition and the person are right. Upload the image to the edit tab and describe only the change you want; a mask lets you repaint one area while the rest stays identical. Regenerate only when the framing or the pose itself is wrong.

Which model gives the best results for photos of people?

For identity, a model that accepts reference images matters more than the model name. Nano Banana 2.1 takes up to 14 references at 3 to 6 credits, and GPT Image 2.5 up to 16 at the same tiering. Use a clean, evenly lit, front-facing reference and put the thing that must not change into the prompt as an instruction.

Open the image generator and write your first specification prompt — the 10 free signup credits cover a round of cheap drafts at 1K.